AI Paper Atlas

Research map / Reinforcement learning / Offline RL

Transitive RL: Value Learning via Divide and Conquer

Seohong Park, Aditya Oberai, Pranav Atreya, Sergey Levine

ICLR 2026 · Reinforcement learning › Offline RL · approach: Offline RL sample complexity theory

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use divide-and-conquer (transitive) value update instead of 1-step TDoffline goal-conditioned RL value learning · puzzle-4x5 (overall)success rate (%): 19 -> 97Improved
use divide-and-conquer value update instead of Monte Carlo value learningoffline goal-conditioned RL value learning · puzzle-4x6 (overall)success rate (%): 37 -> 51Improved
use divide-and-conquer value update instead of MCoffline goal-conditioned RL value learning · humanoidmaze-giant (overall)success rate (%): 79 -> 79No clear effect
use expectile kappa=0.7 instead of 0.5 (behavioral regression) (ablation)TRL · humanoidmaze-giantsuccess rateImproved
use expectile kappa=0.7 instead of 0.5 (ablation)TRL · humanoidmaze-medium and scene-playsuccess rateNo clear effect
use in-trajectory (behavioral) subgoals instead of random subgoals (ablation)TRL · ablation environmentssuccess rateImproved
add distance-based re-weighting (lambda > 0) (ablation)TRL · ablation environmentssuccess rateImproved
use TRL instead of prior GCRL methods on standard tasksoffline GCRL · OGBench (10 environments, 50 tasks)average success rateImproved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy